AI Acceleration and Edge Computing
Closing the von Neumann gap
Conventional computers keep memory and processing apart, and moving data between them dominates the time and energy of modern workloads — the so-called von Neumann bottleneck. The brain has no such separation, and it computes far more per unit of energy. Closing this gap by building complete systems in which memory and computation coincide is the integration goal that the institute's other research areas ultimately feed into: it is the step that turns devices and primitives into hardware that real applications can run on.
This group builds the systems layer — full compute-in-memory accelerators, together with the hardware–software co-design that makes them usable — rather than isolated devices. In practice, that means mapping demanding, modern workloads such as attention, state-space sequence models, and novel network architectures onto memristive in-memory hardware, and integrating memory arrays, peripheral circuitry, and interfaces into working end-to-end chips. A parallel and equally central thread is precision and deployability: the group develops the quantization and training methods that allow these workloads to run at low numerical precision without losing accuracy — so that a design that is efficient is also genuinely usable.

From devices to deployable systems
The team contributed to a demonstration of real-time raw-signal genomic analysis on fully integrated memristor hardware, published in Nature Computational Science (2025) — a rare example of an emerging-device system carried all the way to a complete, real-world application. Leading its own line of work, the group presented a flexible, energy-efficient compute-in-memory accelerator for Kolmogorov–Arnold Networks (Advanced Intelligent Systems, 2026); because these networks replace fixed activation functions with learnable functions along the network's edges, the result extends in-memory computing to a model class beyond the standard multilayer perceptron. On the deployment side, the group showed how modern structured state-space models can run on memristive in-memory hardware (IMSSA, ISCAS 2025) and introduced quantization-aware training to make that deployment up to 100x more efficient (QS4D, Advanced Intelligent Systems, 2025).
Architectures that compute in memory target exactly the workloads where data movement dominates — AI inference, sequence modelling, and data-centric analysis — and they are especially valuable at the edge, where power budgets are tight. The genomics result shows the payoff concretely: a full memristive system delivering analysis in real time, on raw data, in a genuine application setting. The direction of travel is toward larger, more fully integrated compute nodes, and toward rigorous benchmarking against conventional processors on representative end-to-end workloads, so that the efficiency advantage of the whole stack is demonstrated at the system level rather than in isolation.